Pavel Chuprikov, Sergey Nikolenko, Kirill Kogan
Buffering architectures and policies for their efficient management are one of the core ingredients of network architecture. However, despite strong incentives to experiment with and deploy new policies, opportunities for changing or automatically choosing anything beyond a few parameters in a predefined set of behaviors still remain very limited. We introduce a novel buffer management framework based on machine learning approaches which automatically adapts to traffic conditions changing over time and requires only limited knowledge from network operators about the dynamics and optimality of desired behaviors. We validate and compare various design options with a comprehensive evaluation study.